4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Author: Isaac Huang <isaac@clusterfs.com>
40 * 2012-05-13: Liang Zhen <liang@whamcloud.com>
41 * - percpt data for service to improve smp performance
45 #define DEBUG_SUBSYSTEM S_LNET
57 static struct smoketest_rpc {
58 spinlock_t rpc_glock; /* global lock */
59 srpc_service_t *rpc_services[SRPC_SERVICE_MAX_ID + 1];
60 lnet_handle_eq_t rpc_lnet_eq; /* _the_ LNet event queue */
61 srpc_state_t rpc_state;
62 srpc_counters_t rpc_counters;
63 __u64 rpc_matchbits; /* matchbits counter */
67 srpc_serv_portal(int svc_id)
69 return svc_id < SRPC_FRAMEWORK_SERVICE_MAX_ID ?
70 SRPC_FRAMEWORK_REQUEST_PORTAL : SRPC_REQUEST_PORTAL;
74 int srpc_handle_rpc (swi_workitem_t *wi);
76 void srpc_get_counters (srpc_counters_t *cnt)
78 spin_lock(&srpc_data.rpc_glock);
79 *cnt = srpc_data.rpc_counters;
80 spin_unlock(&srpc_data.rpc_glock);
83 void srpc_set_counters (const srpc_counters_t *cnt)
85 spin_lock(&srpc_data.rpc_glock);
86 srpc_data.rpc_counters = *cnt;
87 spin_unlock(&srpc_data.rpc_glock);
91 srpc_add_bulk_page(srpc_bulk_t *bk, struct page *pg, int i, int nob)
93 nob = min(nob, (int)PAGE_CACHE_SIZE);
96 LASSERT(i >= 0 && i < bk->bk_niov);
98 bk->bk_iovs[i].kiov_offset = 0;
99 bk->bk_iovs[i].kiov_page = pg;
100 bk->bk_iovs[i].kiov_len = nob;
105 srpc_free_bulk (srpc_bulk_t *bk)
110 LASSERT (bk != NULL);
112 for (i = 0; i < bk->bk_niov; i++) {
113 pg = bk->bk_iovs[i].kiov_page;
114 if (pg == NULL) break;
119 LIBCFS_FREE(bk, offsetof(srpc_bulk_t, bk_iovs[bk->bk_niov]));
124 srpc_alloc_bulk(int cpt, unsigned bulk_npg, unsigned bulk_len, int sink)
129 LASSERT(bulk_npg > 0 && bulk_npg <= LNET_MAX_IOV);
131 LIBCFS_CPT_ALLOC(bk, lnet_cpt_table(), cpt,
132 offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
134 CERROR("Can't allocate descriptor for %d pages\n", bulk_npg);
138 memset(bk, 0, offsetof(srpc_bulk_t, bk_iovs[bulk_npg]));
140 bk->bk_len = bulk_len;
141 bk->bk_niov = bulk_npg;
143 for (i = 0; i < bulk_npg; i++) {
147 pg = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL);
149 CERROR("Can't allocate page %d of %d\n", i, bulk_npg);
154 nob = srpc_add_bulk_page(bk, pg, i, bulk_len);
166 spin_lock(&srpc_data.rpc_glock);
167 id = srpc_data.rpc_matchbits++;
168 spin_unlock(&srpc_data.rpc_glock);
173 srpc_init_server_rpc(struct srpc_server_rpc *rpc,
174 struct srpc_service_cd *scd,
175 struct srpc_buffer *buffer)
177 memset(rpc, 0, sizeof(*rpc));
178 swi_init_workitem(&rpc->srpc_wi, rpc, srpc_handle_rpc,
179 srpc_serv_is_framework(scd->scd_svc) ?
180 lst_sched_serial : lst_sched_test[scd->scd_cpt]);
182 rpc->srpc_ev.ev_fired = 1; /* no event expected now */
185 rpc->srpc_reqstbuf = buffer;
186 rpc->srpc_peer = buffer->buf_peer;
187 rpc->srpc_self = buffer->buf_self;
188 LNetInvalidateHandle(&rpc->srpc_replymdh);
192 srpc_service_fini(struct srpc_service *svc)
194 struct srpc_service_cd *scd;
195 struct srpc_server_rpc *rpc;
196 struct srpc_buffer *buf;
200 if (svc->sv_cpt_data == NULL)
203 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
205 if (!list_empty(&scd->scd_buf_posted))
206 q = &scd->scd_buf_posted;
207 else if (!list_empty(&scd->scd_buf_blocked))
208 q = &scd->scd_buf_blocked;
212 while (!list_empty(q)) {
213 buf = list_entry(q->next,
216 list_del(&buf->buf_list);
217 LIBCFS_FREE(buf, sizeof(*buf));
221 LASSERT(list_empty(&scd->scd_rpc_active));
223 while (!list_empty(&scd->scd_rpc_free)) {
224 rpc = list_entry(scd->scd_rpc_free.next,
225 struct srpc_server_rpc,
227 list_del(&rpc->srpc_list);
228 LIBCFS_FREE(rpc, sizeof(*rpc));
232 cfs_percpt_free(svc->sv_cpt_data);
233 svc->sv_cpt_data = NULL;
237 srpc_service_nrpcs(struct srpc_service *svc)
239 int nrpcs = svc->sv_wi_total / svc->sv_ncpts;
241 return srpc_serv_is_framework(svc) ?
242 max(nrpcs, SFW_FRWK_WI_MIN) : max(nrpcs, SFW_TEST_WI_MIN);
245 int srpc_add_buffer(struct swi_workitem *wi);
248 srpc_service_init(struct srpc_service *svc)
250 struct srpc_service_cd *scd;
251 struct srpc_server_rpc *rpc;
256 svc->sv_shuttingdown = 0;
258 svc->sv_cpt_data = cfs_percpt_alloc(lnet_cpt_table(),
259 sizeof(struct srpc_service_cd));
260 if (svc->sv_cpt_data == NULL)
263 svc->sv_ncpts = srpc_serv_is_framework(svc) ?
264 1 : cfs_cpt_number(lnet_cpt_table());
265 nrpcs = srpc_service_nrpcs(svc);
267 cfs_percpt_for_each(scd, i, svc->sv_cpt_data) {
270 spin_lock_init(&scd->scd_lock);
271 INIT_LIST_HEAD(&scd->scd_rpc_free);
272 INIT_LIST_HEAD(&scd->scd_rpc_active);
273 INIT_LIST_HEAD(&scd->scd_buf_posted);
274 INIT_LIST_HEAD(&scd->scd_buf_blocked);
276 scd->scd_ev.ev_data = scd;
277 scd->scd_ev.ev_type = SRPC_REQUEST_RCVD;
279 /* NB: don't use lst_sched_serial for adding buffer,
280 * see details in srpc_service_add_buffers() */
281 swi_init_workitem(&scd->scd_buf_wi, scd,
282 srpc_add_buffer, lst_sched_test[i]);
284 if (i != 0 && srpc_serv_is_framework(svc)) {
285 /* NB: framework service only needs srpc_service_cd for
286 * one partition, but we allocate for all to make
287 * it easier to implement, it will waste a little
288 * memory but nobody should care about this */
292 for (j = 0; j < nrpcs; j++) {
293 LIBCFS_CPT_ALLOC(rpc, lnet_cpt_table(),
296 srpc_service_fini(svc);
299 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
307 srpc_add_service(struct srpc_service *sv)
311 LASSERT(0 <= id && id <= SRPC_SERVICE_MAX_ID);
313 if (srpc_service_init(sv) != 0)
316 spin_lock(&srpc_data.rpc_glock);
318 LASSERT(srpc_data.rpc_state == SRPC_STATE_RUNNING);
320 if (srpc_data.rpc_services[id] != NULL) {
321 spin_unlock(&srpc_data.rpc_glock);
325 srpc_data.rpc_services[id] = sv;
326 spin_unlock(&srpc_data.rpc_glock);
328 CDEBUG(D_NET, "Adding service: id %d, name %s\n", id, sv->sv_name);
332 srpc_service_fini(sv);
337 srpc_remove_service (srpc_service_t *sv)
341 spin_lock(&srpc_data.rpc_glock);
343 if (srpc_data.rpc_services[id] != sv) {
344 spin_unlock(&srpc_data.rpc_glock);
348 srpc_data.rpc_services[id] = NULL;
349 spin_unlock(&srpc_data.rpc_glock);
354 srpc_post_passive_rdma(int portal, int local, __u64 matchbits, void *buf,
355 int len, int options, lnet_process_id_t peer,
356 lnet_handle_md_t *mdh, srpc_event_t *ev)
360 lnet_handle_me_t meh;
362 rc = LNetMEAttach(portal, peer, matchbits, 0, LNET_UNLINK,
363 local ? LNET_INS_LOCAL : LNET_INS_AFTER, &meh);
365 CERROR ("LNetMEAttach failed: %d\n", rc);
366 LASSERT (rc == -ENOMEM);
374 md.options = options;
375 md.eq_handle = srpc_data.rpc_lnet_eq;
377 rc = LNetMDAttach(meh, md, LNET_UNLINK, mdh);
379 CERROR ("LNetMDAttach failed: %d\n", rc);
380 LASSERT (rc == -ENOMEM);
382 rc = LNetMEUnlink(meh);
388 "Posted passive RDMA: peer %s, portal %d, matchbits %#llx\n",
389 libcfs_id2str(peer), portal, matchbits);
394 srpc_post_active_rdma(int portal, __u64 matchbits, void *buf, int len,
395 int options, lnet_process_id_t peer, lnet_nid_t self,
396 lnet_handle_md_t *mdh, srpc_event_t *ev)
404 md.eq_handle = srpc_data.rpc_lnet_eq;
405 md.threshold = ((options & LNET_MD_OP_GET) != 0) ? 2 : 1;
406 md.options = options & ~(LNET_MD_OP_PUT | LNET_MD_OP_GET);
408 rc = LNetMDBind(md, LNET_UNLINK, mdh);
410 CERROR ("LNetMDBind failed: %d\n", rc);
411 LASSERT (rc == -ENOMEM);
415 /* this is kind of an abuse of the LNET_MD_OP_{PUT,GET} options.
416 * they're only meaningful for MDs attached to an ME (i.e. passive
418 if ((options & LNET_MD_OP_PUT) != 0) {
419 rc = LNetPut(self, *mdh, LNET_NOACK_REQ, peer,
420 portal, matchbits, 0, 0);
422 LASSERT ((options & LNET_MD_OP_GET) != 0);
424 rc = LNetGet(self, *mdh, peer, portal, matchbits, 0);
428 CERROR ("LNet%s(%s, %d, %lld) failed: %d\n",
429 ((options & LNET_MD_OP_PUT) != 0) ? "Put" : "Get",
430 libcfs_id2str(peer), portal, matchbits, rc);
432 /* The forthcoming unlink event will complete this operation
433 * with failure, so fall through and return success here.
435 rc = LNetMDUnlink(*mdh);
439 "Posted active RDMA: peer %s, portal %u, matchbits %#llx\n",
440 libcfs_id2str(peer), portal, matchbits);
446 srpc_post_passive_rqtbuf(int service, int local, void *buf, int len,
447 lnet_handle_md_t *mdh, srpc_event_t *ev)
449 lnet_process_id_t any = {0};
451 any.nid = LNET_NID_ANY;
452 any.pid = LNET_PID_ANY;
454 return srpc_post_passive_rdma(srpc_serv_portal(service),
455 local, service, buf, len,
456 LNET_MD_OP_PUT, any, mdh, ev);
460 srpc_service_post_buffer(struct srpc_service_cd *scd, struct srpc_buffer *buf)
461 __must_hold(&scd->scd_lock)
463 struct srpc_service *sv = scd->scd_svc;
464 struct srpc_msg *msg = &buf->buf_msg;
467 LNetInvalidateHandle(&buf->buf_mdh);
468 list_add(&buf->buf_list, &scd->scd_buf_posted);
469 scd->scd_buf_nposted++;
470 spin_unlock(&scd->scd_lock);
472 rc = srpc_post_passive_rqtbuf(sv->sv_id,
473 !srpc_serv_is_framework(sv),
474 msg, sizeof(*msg), &buf->buf_mdh,
477 /* At this point, a RPC (new or delayed) may have arrived in
478 * msg and its event handler has been called. So we must add
479 * buf to scd_buf_posted _before_ dropping scd_lock */
481 spin_lock(&scd->scd_lock);
484 if (!sv->sv_shuttingdown)
487 spin_unlock(&scd->scd_lock);
488 /* srpc_shutdown_service might have tried to unlink me
489 * when my buf_mdh was still invalid */
490 LNetMDUnlink(buf->buf_mdh);
491 spin_lock(&scd->scd_lock);
495 scd->scd_buf_nposted--;
496 if (sv->sv_shuttingdown)
497 return rc; /* don't allow to change scd_buf_posted */
499 list_del(&buf->buf_list);
500 spin_unlock(&scd->scd_lock);
502 LIBCFS_FREE(buf, sizeof(*buf));
504 spin_lock(&scd->scd_lock);
509 srpc_add_buffer(struct swi_workitem *wi)
511 struct srpc_service_cd *scd = wi->swi_workitem.wi_data;
512 struct srpc_buffer *buf;
515 /* it's called by workitem scheduler threads, these threads
516 * should have been set CPT affinity, so buffers will be posted
517 * on CPT local list of Portal */
518 spin_lock(&scd->scd_lock);
520 while (scd->scd_buf_adjust > 0 &&
521 !scd->scd_svc->sv_shuttingdown) {
522 scd->scd_buf_adjust--; /* consume it */
523 scd->scd_buf_posting++;
525 spin_unlock(&scd->scd_lock);
527 LIBCFS_ALLOC(buf, sizeof(*buf));
529 CERROR("Failed to add new buf to service: %s\n",
530 scd->scd_svc->sv_name);
531 spin_lock(&scd->scd_lock);
536 spin_lock(&scd->scd_lock);
537 if (scd->scd_svc->sv_shuttingdown) {
538 spin_unlock(&scd->scd_lock);
539 LIBCFS_FREE(buf, sizeof(*buf));
541 spin_lock(&scd->scd_lock);
546 rc = srpc_service_post_buffer(scd, buf);
548 break; /* buf has been freed inside */
550 LASSERT(scd->scd_buf_posting > 0);
551 scd->scd_buf_posting--;
552 scd->scd_buf_total++;
553 scd->scd_buf_low = MAX(2, scd->scd_buf_total / 4);
557 scd->scd_buf_err_stamp = cfs_time_current_sec();
558 scd->scd_buf_err = rc;
560 LASSERT(scd->scd_buf_posting > 0);
561 scd->scd_buf_posting--;
564 spin_unlock(&scd->scd_lock);
569 srpc_service_add_buffers(struct srpc_service *sv, int nbuffer)
571 struct srpc_service_cd *scd;
575 LASSERTF(nbuffer > 0, "nbuffer must be positive: %d\n", nbuffer);
577 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
578 spin_lock(&scd->scd_lock);
580 scd->scd_buf_err = 0;
581 scd->scd_buf_err_stamp = 0;
582 scd->scd_buf_posting = 0;
583 scd->scd_buf_adjust = nbuffer;
584 /* start to post buffers */
585 swi_schedule_workitem(&scd->scd_buf_wi);
586 spin_unlock(&scd->scd_lock);
588 /* framework service only post buffer for one partition */
589 if (srpc_serv_is_framework(sv))
593 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
594 spin_lock(&scd->scd_lock);
596 * NB: srpc_service_add_buffers() can be called inside
597 * thread context of lst_sched_serial, and we don't normally
598 * allow to sleep inside thread context of WI scheduler
599 * because it will block current scheduler thread from doing
600 * anything else, even worse, it could deadlock if it's
601 * waiting on result from another WI of the same scheduler.
602 * However, it's safe at here because scd_buf_wi is scheduled
603 * by thread in a different WI scheduler (lst_sched_test),
604 * so we don't have any risk of deadlock, though this could
605 * block all WIs pending on lst_sched_serial for a moment
606 * which is not good but not fatal.
608 lst_wait_until(scd->scd_buf_err != 0 ||
609 (scd->scd_buf_adjust == 0 &&
610 scd->scd_buf_posting == 0),
611 scd->scd_lock, "waiting for adding buffer\n");
613 if (scd->scd_buf_err != 0 && rc == 0)
614 rc = scd->scd_buf_err;
616 spin_unlock(&scd->scd_lock);
623 srpc_service_remove_buffers(struct srpc_service *sv, int nbuffer)
625 struct srpc_service_cd *scd;
629 LASSERT(!sv->sv_shuttingdown);
631 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
632 spin_lock(&scd->scd_lock);
634 num = scd->scd_buf_total + scd->scd_buf_posting;
635 scd->scd_buf_adjust -= min(nbuffer, num);
637 spin_unlock(&scd->scd_lock);
641 /* returns 1 if sv has finished, otherwise 0 */
643 srpc_finish_service(struct srpc_service *sv)
645 struct srpc_service_cd *scd;
646 struct srpc_server_rpc *rpc;
649 LASSERT(sv->sv_shuttingdown); /* srpc_shutdown_service called */
651 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
652 spin_lock(&scd->scd_lock);
653 if (!swi_deschedule_workitem(&scd->scd_buf_wi)) {
654 spin_unlock(&scd->scd_lock);
658 if (scd->scd_buf_nposted > 0) {
659 CDEBUG(D_NET, "waiting for %d posted buffers to "
660 "unlink\n", scd->scd_buf_nposted);
661 spin_unlock(&scd->scd_lock);
665 if (list_empty(&scd->scd_rpc_active)) {
666 spin_unlock(&scd->scd_lock);
670 rpc = list_entry(scd->scd_rpc_active.next,
671 struct srpc_server_rpc, srpc_list);
672 CNETERR("Active RPC %p on shutdown: sv %s, peer %s, "
673 "wi %s scheduled %d running %d, "
674 "ev fired %d type %d status %d lnet %d\n",
675 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
676 swi_state2str(rpc->srpc_wi.swi_state),
677 rpc->srpc_wi.swi_workitem.wi_scheduled,
678 rpc->srpc_wi.swi_workitem.wi_running,
679 rpc->srpc_ev.ev_fired, rpc->srpc_ev.ev_type,
680 rpc->srpc_ev.ev_status, rpc->srpc_ev.ev_lnet);
681 spin_unlock(&scd->scd_lock);
685 /* no lock needed from now on */
686 srpc_service_fini(sv);
690 /* called with sv->sv_lock held */
692 srpc_service_recycle_buffer(struct srpc_service_cd *scd, srpc_buffer_t *buf)
693 __must_hold(&scd->scd_lock)
695 if (!scd->scd_svc->sv_shuttingdown && scd->scd_buf_adjust >= 0) {
696 if (srpc_service_post_buffer(scd, buf) != 0) {
697 CWARN("Failed to post %s buffer\n",
698 scd->scd_svc->sv_name);
703 /* service is shutting down, or we want to recycle some buffers */
704 scd->scd_buf_total--;
706 if (scd->scd_buf_adjust < 0) {
707 scd->scd_buf_adjust++;
708 if (scd->scd_buf_adjust < 0 &&
709 scd->scd_buf_total == 0 && scd->scd_buf_posting == 0) {
711 "Try to recyle %d buffers but nothing left\n",
712 scd->scd_buf_adjust);
713 scd->scd_buf_adjust = 0;
717 spin_unlock(&scd->scd_lock);
718 LIBCFS_FREE(buf, sizeof(*buf));
719 spin_lock(&scd->scd_lock);
723 srpc_abort_service(struct srpc_service *sv)
725 struct srpc_service_cd *scd;
726 struct srpc_server_rpc *rpc;
729 CDEBUG(D_NET, "Aborting service: id %d, name %s\n",
730 sv->sv_id, sv->sv_name);
732 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
733 spin_lock(&scd->scd_lock);
735 /* schedule in-flight RPCs to notice the abort, NB:
736 * racing with incoming RPCs; complete fix should make test
737 * RPCs carry session ID in its headers */
738 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list) {
739 rpc->srpc_aborted = 1;
740 swi_schedule_workitem(&rpc->srpc_wi);
743 spin_unlock(&scd->scd_lock);
748 srpc_shutdown_service(srpc_service_t *sv)
750 struct srpc_service_cd *scd;
751 struct srpc_server_rpc *rpc;
755 CDEBUG(D_NET, "Shutting down service: id %d, name %s\n",
756 sv->sv_id, sv->sv_name);
758 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
759 spin_lock(&scd->scd_lock);
761 sv->sv_shuttingdown = 1; /* i.e. no new active RPC */
763 cfs_percpt_for_each(scd, i, sv->sv_cpt_data)
764 spin_unlock(&scd->scd_lock);
766 cfs_percpt_for_each(scd, i, sv->sv_cpt_data) {
767 spin_lock(&scd->scd_lock);
769 /* schedule in-flight RPCs to notice the shutdown */
770 list_for_each_entry(rpc, &scd->scd_rpc_active, srpc_list)
771 swi_schedule_workitem(&rpc->srpc_wi);
773 spin_unlock(&scd->scd_lock);
775 /* OK to traverse scd_buf_posted without lock, since no one
776 * touches scd_buf_posted now */
777 list_for_each_entry(buf, &scd->scd_buf_posted, buf_list)
778 LNetMDUnlink(buf->buf_mdh);
783 srpc_send_request (srpc_client_rpc_t *rpc)
785 srpc_event_t *ev = &rpc->crpc_reqstev;
790 ev->ev_type = SRPC_REQUEST_SENT;
792 rc = srpc_post_active_rdma(srpc_serv_portal(rpc->crpc_service),
793 rpc->crpc_service, &rpc->crpc_reqstmsg,
794 sizeof(srpc_msg_t), LNET_MD_OP_PUT,
795 rpc->crpc_dest, LNET_NID_ANY,
796 &rpc->crpc_reqstmdh, ev);
798 LASSERT (rc == -ENOMEM);
799 ev->ev_fired = 1; /* no more event expected */
805 srpc_prepare_reply (srpc_client_rpc_t *rpc)
807 srpc_event_t *ev = &rpc->crpc_replyev;
808 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.rpyid;
813 ev->ev_type = SRPC_REPLY_RCVD;
815 *id = srpc_next_id();
817 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
818 &rpc->crpc_replymsg, sizeof(srpc_msg_t),
819 LNET_MD_OP_PUT, rpc->crpc_dest,
820 &rpc->crpc_replymdh, ev);
822 LASSERT (rc == -ENOMEM);
823 ev->ev_fired = 1; /* no more event expected */
829 srpc_prepare_bulk (srpc_client_rpc_t *rpc)
831 srpc_bulk_t *bk = &rpc->crpc_bulk;
832 srpc_event_t *ev = &rpc->crpc_bulkev;
833 __u64 *id = &rpc->crpc_reqstmsg.msg_body.reqst.bulkid;
837 LASSERT (bk->bk_niov <= LNET_MAX_IOV);
839 if (bk->bk_niov == 0) return 0; /* nothing to do */
841 opt = bk->bk_sink ? LNET_MD_OP_PUT : LNET_MD_OP_GET;
846 ev->ev_type = SRPC_BULK_REQ_RCVD;
848 *id = srpc_next_id();
850 rc = srpc_post_passive_rdma(SRPC_RDMA_PORTAL, 0, *id,
851 &bk->bk_iovs[0], bk->bk_niov, opt,
852 rpc->crpc_dest, &bk->bk_mdh, ev);
854 LASSERT (rc == -ENOMEM);
855 ev->ev_fired = 1; /* no more event expected */
861 srpc_do_bulk (srpc_server_rpc_t *rpc)
863 srpc_event_t *ev = &rpc->srpc_ev;
864 srpc_bulk_t *bk = rpc->srpc_bulk;
865 __u64 id = rpc->srpc_reqstbuf->buf_msg.msg_body.reqst.bulkid;
869 LASSERT (bk != NULL);
871 opt = bk->bk_sink ? LNET_MD_OP_GET : LNET_MD_OP_PUT;
876 ev->ev_type = bk->bk_sink ? SRPC_BULK_GET_RPLD : SRPC_BULK_PUT_SENT;
878 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, id,
879 &bk->bk_iovs[0], bk->bk_niov, opt,
880 rpc->srpc_peer, rpc->srpc_self,
883 ev->ev_fired = 1; /* no more event expected */
887 /* only called from srpc_handle_rpc */
889 srpc_server_rpc_done(srpc_server_rpc_t *rpc, int status)
891 struct srpc_service_cd *scd = rpc->srpc_scd;
892 struct srpc_service *sv = scd->scd_svc;
893 srpc_buffer_t *buffer;
895 LASSERT (status != 0 || rpc->srpc_wi.swi_state == SWI_STATE_DONE);
897 rpc->srpc_status = status;
899 CDEBUG_LIMIT (status == 0 ? D_NET : D_NETERROR,
900 "Server RPC %p done: service %s, peer %s, status %s:%d\n",
901 rpc, sv->sv_name, libcfs_id2str(rpc->srpc_peer),
902 swi_state2str(rpc->srpc_wi.swi_state), status);
905 spin_lock(&srpc_data.rpc_glock);
906 srpc_data.rpc_counters.rpcs_dropped++;
907 spin_unlock(&srpc_data.rpc_glock);
910 if (rpc->srpc_done != NULL)
911 (*rpc->srpc_done) (rpc);
912 LASSERT(rpc->srpc_bulk == NULL);
914 spin_lock(&scd->scd_lock);
916 if (rpc->srpc_reqstbuf != NULL) {
917 /* NB might drop sv_lock in srpc_service_recycle_buffer, but
918 * sv won't go away for scd_rpc_active must not be empty */
919 srpc_service_recycle_buffer(scd, rpc->srpc_reqstbuf);
920 rpc->srpc_reqstbuf = NULL;
923 list_del(&rpc->srpc_list); /* from scd->scd_rpc_active */
926 * No one can schedule me now since:
927 * - I'm not on scd_rpc_active.
928 * - all LNet events have been fired.
929 * Cancel pending schedules and prevent future schedule attempts:
931 LASSERT(rpc->srpc_ev.ev_fired);
932 swi_exit_workitem(&rpc->srpc_wi);
934 if (!sv->sv_shuttingdown && !list_empty(&scd->scd_buf_blocked)) {
935 buffer = list_entry(scd->scd_buf_blocked.next,
936 srpc_buffer_t, buf_list);
937 list_del(&buffer->buf_list);
939 srpc_init_server_rpc(rpc, scd, buffer);
940 list_add_tail(&rpc->srpc_list, &scd->scd_rpc_active);
941 swi_schedule_workitem(&rpc->srpc_wi);
943 list_add(&rpc->srpc_list, &scd->scd_rpc_free);
946 spin_unlock(&scd->scd_lock);
950 /* handles an incoming RPC */
952 srpc_handle_rpc(swi_workitem_t *wi)
954 struct srpc_server_rpc *rpc = wi->swi_workitem.wi_data;
955 struct srpc_service_cd *scd = rpc->srpc_scd;
956 struct srpc_service *sv = scd->scd_svc;
957 srpc_event_t *ev = &rpc->srpc_ev;
960 LASSERT(wi == &rpc->srpc_wi);
962 spin_lock(&scd->scd_lock);
964 if (sv->sv_shuttingdown || rpc->srpc_aborted) {
965 spin_unlock(&scd->scd_lock);
967 if (rpc->srpc_bulk != NULL)
968 LNetMDUnlink(rpc->srpc_bulk->bk_mdh);
969 LNetMDUnlink(rpc->srpc_replymdh);
971 if (ev->ev_fired) { /* no more event, OK to finish */
972 srpc_server_rpc_done(rpc, -ESHUTDOWN);
978 spin_unlock(&scd->scd_lock);
980 switch (wi->swi_state) {
983 case SWI_STATE_NEWBORN: {
985 srpc_generic_reply_t *reply;
987 msg = &rpc->srpc_reqstbuf->buf_msg;
988 reply = &rpc->srpc_replymsg.msg_body.reply;
990 if (msg->msg_magic == 0) {
991 /* moaned already in srpc_lnet_ev_handler */
992 srpc_server_rpc_done(rpc, EBADMSG);
996 srpc_unpack_msg_hdr(msg);
997 if (msg->msg_version != SRPC_MSG_VERSION) {
998 CWARN("Version mismatch: %u, %u expected, from %s\n",
999 msg->msg_version, SRPC_MSG_VERSION,
1000 libcfs_id2str(rpc->srpc_peer));
1001 reply->status = EPROTO;
1002 /* drop through and send reply */
1005 rc = (*sv->sv_handler)(rpc);
1006 LASSERT(reply->status == 0 || !rpc->srpc_bulk);
1008 srpc_server_rpc_done(rpc, rc);
1013 wi->swi_state = SWI_STATE_BULK_STARTED;
1015 if (rpc->srpc_bulk != NULL) {
1016 rc = srpc_do_bulk(rpc);
1018 return 0; /* wait for bulk */
1020 LASSERT (ev->ev_fired);
1024 case SWI_STATE_BULK_STARTED:
1025 LASSERT (rpc->srpc_bulk == NULL || ev->ev_fired);
1027 if (rpc->srpc_bulk != NULL) {
1030 if (sv->sv_bulk_ready != NULL)
1031 rc = (*sv->sv_bulk_ready) (rpc, rc);
1034 srpc_server_rpc_done(rpc, rc);
1039 wi->swi_state = SWI_STATE_REPLY_SUBMITTED;
1040 rc = srpc_send_reply(rpc);
1042 return 0; /* wait for reply */
1043 srpc_server_rpc_done(rpc, rc);
1046 case SWI_STATE_REPLY_SUBMITTED:
1047 if (!ev->ev_fired) {
1048 CERROR("RPC %p: bulk %p, service %d\n",
1049 rpc, rpc->srpc_bulk, sv->sv_id);
1050 CERROR("Event: status %d, type %d, lnet %d\n",
1051 ev->ev_status, ev->ev_type, ev->ev_lnet);
1052 LASSERT (ev->ev_fired);
1055 wi->swi_state = SWI_STATE_DONE;
1056 srpc_server_rpc_done(rpc, ev->ev_status);
1064 srpc_client_rpc_expired (void *data)
1066 srpc_client_rpc_t *rpc = data;
1068 CWARN ("Client RPC expired: service %d, peer %s, timeout %d.\n",
1069 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1072 spin_lock(&rpc->crpc_lock);
1074 rpc->crpc_timeout = 0;
1075 srpc_abort_rpc(rpc, -ETIMEDOUT);
1077 spin_unlock(&rpc->crpc_lock);
1079 spin_lock(&srpc_data.rpc_glock);
1080 srpc_data.rpc_counters.rpcs_expired++;
1081 spin_unlock(&srpc_data.rpc_glock);
1085 srpc_add_client_rpc_timer(srpc_client_rpc_t *rpc)
1087 stt_timer_t *timer = &rpc->crpc_timer;
1089 if (rpc->crpc_timeout == 0)
1092 INIT_LIST_HEAD(&timer->stt_list);
1093 timer->stt_data = rpc;
1094 timer->stt_func = srpc_client_rpc_expired;
1095 timer->stt_expires = cfs_time_add(rpc->crpc_timeout,
1096 cfs_time_current_sec());
1097 stt_add_timer(timer);
1102 * Called with rpc->crpc_lock held.
1104 * Upon exit the RPC expiry timer is not queued and the handler is not
1105 * running on any CPU. */
1107 srpc_del_client_rpc_timer (srpc_client_rpc_t *rpc)
1109 /* timer not planted or already exploded */
1110 if (rpc->crpc_timeout == 0)
1113 /* timer successfully defused */
1114 if (stt_del_timer(&rpc->crpc_timer))
1117 /* timer detonated, wait for it to explode */
1118 while (rpc->crpc_timeout != 0) {
1119 spin_unlock(&rpc->crpc_lock);
1123 spin_lock(&rpc->crpc_lock);
1128 srpc_client_rpc_done (srpc_client_rpc_t *rpc, int status)
1130 swi_workitem_t *wi = &rpc->crpc_wi;
1132 LASSERT(status != 0 || wi->swi_state == SWI_STATE_DONE);
1134 spin_lock(&rpc->crpc_lock);
1136 rpc->crpc_closed = 1;
1137 if (rpc->crpc_status == 0)
1138 rpc->crpc_status = status;
1140 srpc_del_client_rpc_timer(rpc);
1142 CDEBUG_LIMIT ((status == 0) ? D_NET : D_NETERROR,
1143 "Client RPC done: service %d, peer %s, status %s:%d:%d\n",
1144 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1145 swi_state2str(wi->swi_state), rpc->crpc_aborted, status);
1148 * No one can schedule me now since:
1149 * - RPC timer has been defused.
1150 * - all LNet events have been fired.
1151 * - crpc_closed has been set, preventing srpc_abort_rpc from
1153 * Cancel pending schedules and prevent future schedule attempts:
1155 LASSERT (!srpc_event_pending(rpc));
1156 swi_exit_workitem(wi);
1158 spin_unlock(&rpc->crpc_lock);
1160 (*rpc->crpc_done)(rpc);
1164 /* sends an outgoing RPC */
1166 srpc_send_rpc (swi_workitem_t *wi)
1169 srpc_client_rpc_t *rpc;
1173 LASSERT(wi != NULL);
1175 rpc = wi->swi_workitem.wi_data;
1177 LASSERT (rpc != NULL);
1178 LASSERT (wi == &rpc->crpc_wi);
1180 reply = &rpc->crpc_replymsg;
1181 do_bulk = rpc->crpc_bulk.bk_niov > 0;
1183 spin_lock(&rpc->crpc_lock);
1185 if (rpc->crpc_aborted) {
1186 spin_unlock(&rpc->crpc_lock);
1190 spin_unlock(&rpc->crpc_lock);
1192 switch (wi->swi_state) {
1195 case SWI_STATE_NEWBORN:
1196 LASSERT (!srpc_event_pending(rpc));
1198 rc = srpc_prepare_reply(rpc);
1200 srpc_client_rpc_done(rpc, rc);
1204 rc = srpc_prepare_bulk(rpc);
1207 wi->swi_state = SWI_STATE_REQUEST_SUBMITTED;
1208 rc = srpc_send_request(rpc);
1211 case SWI_STATE_REQUEST_SUBMITTED:
1212 /* CAVEAT EMPTOR: rqtev, rpyev, and bulkev may come in any
1213 * order; however, they're processed in a strict order:
1214 * rqt, rpy, and bulk. */
1215 if (!rpc->crpc_reqstev.ev_fired) break;
1217 rc = rpc->crpc_reqstev.ev_status;
1220 wi->swi_state = SWI_STATE_REQUEST_SENT;
1221 /* perhaps more events, fall thru */
1222 case SWI_STATE_REQUEST_SENT: {
1223 srpc_msg_type_t type = srpc_service2reply(rpc->crpc_service);
1225 if (!rpc->crpc_replyev.ev_fired) break;
1227 rc = rpc->crpc_replyev.ev_status;
1230 srpc_unpack_msg_hdr(reply);
1231 if (reply->msg_type != type ||
1232 (reply->msg_magic != SRPC_MSG_MAGIC &&
1233 reply->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1234 CWARN ("Bad message from %s: type %u (%d expected),"
1235 " magic %u (%d expected).\n",
1236 libcfs_id2str(rpc->crpc_dest),
1237 reply->msg_type, type,
1238 reply->msg_magic, SRPC_MSG_MAGIC);
1243 if (do_bulk && reply->msg_body.reply.status != 0) {
1244 CWARN ("Remote error %d at %s, unlink bulk buffer in "
1245 "case peer didn't initiate bulk transfer\n",
1246 reply->msg_body.reply.status,
1247 libcfs_id2str(rpc->crpc_dest));
1248 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1251 wi->swi_state = SWI_STATE_REPLY_RECEIVED;
1253 case SWI_STATE_REPLY_RECEIVED:
1254 if (do_bulk && !rpc->crpc_bulkev.ev_fired) break;
1256 rc = do_bulk ? rpc->crpc_bulkev.ev_status : 0;
1258 /* Bulk buffer was unlinked due to remote error. Clear error
1259 * since reply buffer still contains valid data.
1260 * NB rpc->crpc_done shouldn't look into bulk data in case of
1262 if (do_bulk && rpc->crpc_bulkev.ev_lnet == LNET_EVENT_UNLINK &&
1263 rpc->crpc_status == 0 && reply->msg_body.reply.status != 0)
1266 wi->swi_state = SWI_STATE_DONE;
1267 srpc_client_rpc_done(rpc, rc);
1272 spin_lock(&rpc->crpc_lock);
1273 srpc_abort_rpc(rpc, rc);
1274 spin_unlock(&rpc->crpc_lock);
1278 if (rpc->crpc_aborted) {
1279 LNetMDUnlink(rpc->crpc_reqstmdh);
1280 LNetMDUnlink(rpc->crpc_replymdh);
1281 LNetMDUnlink(rpc->crpc_bulk.bk_mdh);
1283 if (!srpc_event_pending(rpc)) {
1284 srpc_client_rpc_done(rpc, -EINTR);
1292 srpc_create_client_rpc (lnet_process_id_t peer, int service,
1293 int nbulkiov, int bulklen,
1294 void (*rpc_done)(srpc_client_rpc_t *),
1295 void (*rpc_fini)(srpc_client_rpc_t *), void *priv)
1297 srpc_client_rpc_t *rpc;
1299 LIBCFS_ALLOC(rpc, offsetof(srpc_client_rpc_t,
1300 crpc_bulk.bk_iovs[nbulkiov]));
1304 srpc_init_client_rpc(rpc, peer, service, nbulkiov,
1305 bulklen, rpc_done, rpc_fini, priv);
1309 /* called with rpc->crpc_lock held */
1311 srpc_abort_rpc (srpc_client_rpc_t *rpc, int why)
1315 if (rpc->crpc_aborted || /* already aborted */
1316 rpc->crpc_closed) /* callback imminent */
1320 "Aborting RPC: service %d, peer %s, state %s, why %d\n",
1321 rpc->crpc_service, libcfs_id2str(rpc->crpc_dest),
1322 swi_state2str(rpc->crpc_wi.swi_state), why);
1324 rpc->crpc_aborted = 1;
1325 rpc->crpc_status = why;
1326 swi_schedule_workitem(&rpc->crpc_wi);
1330 /* called with rpc->crpc_lock held */
1332 srpc_post_rpc (srpc_client_rpc_t *rpc)
1334 LASSERT (!rpc->crpc_aborted);
1335 LASSERT (srpc_data.rpc_state == SRPC_STATE_RUNNING);
1337 CDEBUG (D_NET, "Posting RPC: peer %s, service %d, timeout %d\n",
1338 libcfs_id2str(rpc->crpc_dest), rpc->crpc_service,
1341 srpc_add_client_rpc_timer(rpc);
1342 swi_schedule_workitem(&rpc->crpc_wi);
1348 srpc_send_reply(struct srpc_server_rpc *rpc)
1350 srpc_event_t *ev = &rpc->srpc_ev;
1351 struct srpc_msg *msg = &rpc->srpc_replymsg;
1352 struct srpc_buffer *buffer = rpc->srpc_reqstbuf;
1353 struct srpc_service_cd *scd = rpc->srpc_scd;
1354 struct srpc_service *sv = scd->scd_svc;
1358 LASSERT(buffer != NULL);
1359 rpyid = buffer->buf_msg.msg_body.reqst.rpyid;
1361 spin_lock(&scd->scd_lock);
1363 if (!sv->sv_shuttingdown && !srpc_serv_is_framework(sv)) {
1364 /* Repost buffer before replying since test client
1365 * might send me another RPC once it gets the reply */
1366 if (srpc_service_post_buffer(scd, buffer) != 0)
1367 CWARN("Failed to repost %s buffer\n", sv->sv_name);
1368 rpc->srpc_reqstbuf = NULL;
1371 spin_unlock(&scd->scd_lock);
1375 ev->ev_type = SRPC_REPLY_SENT;
1377 msg->msg_magic = SRPC_MSG_MAGIC;
1378 msg->msg_version = SRPC_MSG_VERSION;
1379 msg->msg_type = srpc_service2reply(sv->sv_id);
1381 rc = srpc_post_active_rdma(SRPC_RDMA_PORTAL, rpyid, msg,
1382 sizeof(*msg), LNET_MD_OP_PUT,
1383 rpc->srpc_peer, rpc->srpc_self,
1384 &rpc->srpc_replymdh, ev);
1386 ev->ev_fired = 1; /* no more event expected */
1390 /* when in kernel always called with LNET_LOCK() held, and in thread context */
1392 srpc_lnet_ev_handler(lnet_event_t *ev)
1394 struct srpc_service_cd *scd;
1395 srpc_event_t *rpcev = ev->md.user_ptr;
1396 srpc_client_rpc_t *crpc;
1397 srpc_server_rpc_t *srpc;
1398 srpc_buffer_t *buffer;
1401 srpc_msg_type_t type;
1403 LASSERT (!in_interrupt());
1405 if (ev->status != 0) {
1408 spin_lock(&srpc_data.rpc_glock);
1409 if (ev->status != -ECANCELED) /* cancellation is not error */
1410 srpc_data.rpc_counters.errors++;
1411 errors = srpc_data.rpc_counters.errors;
1412 spin_unlock(&srpc_data.rpc_glock);
1414 CNETERR("LNet event status %d type %d, RPC errors %u\n",
1415 ev->status, ev->type, errors);
1418 rpcev->ev_lnet = ev->type;
1420 switch (rpcev->ev_type) {
1422 CERROR("Unknown event: status %d, type %d, lnet %d\n",
1423 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1425 case SRPC_REQUEST_SENT:
1426 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1427 spin_lock(&srpc_data.rpc_glock);
1428 srpc_data.rpc_counters.rpcs_sent++;
1429 spin_unlock(&srpc_data.rpc_glock);
1431 case SRPC_REPLY_RCVD:
1432 case SRPC_BULK_REQ_RCVD:
1433 crpc = rpcev->ev_data;
1435 if (rpcev != &crpc->crpc_reqstev &&
1436 rpcev != &crpc->crpc_replyev &&
1437 rpcev != &crpc->crpc_bulkev) {
1438 CERROR("rpcev %p, crpc %p, reqstev %p, replyev %p, bulkev %p\n",
1439 rpcev, crpc, &crpc->crpc_reqstev,
1440 &crpc->crpc_replyev, &crpc->crpc_bulkev);
1441 CERROR("Bad event: status %d, type %d, lnet %d\n",
1442 rpcev->ev_status, rpcev->ev_type, rpcev->ev_lnet);
1446 spin_lock(&crpc->crpc_lock);
1448 LASSERT(rpcev->ev_fired == 0);
1449 rpcev->ev_fired = 1;
1450 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1451 -EINTR : ev->status;
1452 swi_schedule_workitem(&crpc->crpc_wi);
1454 spin_unlock(&crpc->crpc_lock);
1457 case SRPC_REQUEST_RCVD:
1458 scd = rpcev->ev_data;
1461 LASSERT(rpcev == &scd->scd_ev);
1463 spin_lock(&scd->scd_lock);
1465 LASSERT (ev->unlinked);
1466 LASSERT (ev->type == LNET_EVENT_PUT ||
1467 ev->type == LNET_EVENT_UNLINK);
1468 LASSERT (ev->type != LNET_EVENT_UNLINK ||
1469 sv->sv_shuttingdown);
1471 buffer = container_of(ev->md.start, srpc_buffer_t, buf_msg);
1472 buffer->buf_peer = ev->initiator;
1473 buffer->buf_self = ev->target.nid;
1475 LASSERT(scd->scd_buf_nposted > 0);
1476 scd->scd_buf_nposted--;
1478 if (sv->sv_shuttingdown) {
1479 /* Leave buffer on scd->scd_buf_nposted since
1480 * srpc_finish_service needs to traverse it. */
1481 spin_unlock(&scd->scd_lock);
1485 if (scd->scd_buf_err_stamp != 0 &&
1486 scd->scd_buf_err_stamp < cfs_time_current_sec()) {
1487 /* re-enable adding buffer */
1488 scd->scd_buf_err_stamp = 0;
1489 scd->scd_buf_err = 0;
1492 if (scd->scd_buf_err == 0 && /* adding buffer is enabled */
1493 scd->scd_buf_adjust == 0 &&
1494 scd->scd_buf_nposted < scd->scd_buf_low) {
1495 scd->scd_buf_adjust = MAX(scd->scd_buf_total / 2,
1497 swi_schedule_workitem(&scd->scd_buf_wi);
1500 list_del(&buffer->buf_list); /* from scd->scd_buf_posted */
1501 msg = &buffer->buf_msg;
1502 type = srpc_service2request(sv->sv_id);
1504 if (ev->status != 0 || ev->mlength != sizeof(*msg) ||
1505 (msg->msg_type != type &&
1506 msg->msg_type != __swab32(type)) ||
1507 (msg->msg_magic != SRPC_MSG_MAGIC &&
1508 msg->msg_magic != __swab32(SRPC_MSG_MAGIC))) {
1509 CERROR ("Dropping RPC (%s) from %s: "
1510 "status %d mlength %d type %u magic %u.\n",
1511 sv->sv_name, libcfs_id2str(ev->initiator),
1512 ev->status, ev->mlength,
1513 msg->msg_type, msg->msg_magic);
1515 /* NB can't call srpc_service_recycle_buffer here since
1516 * it may call LNetM[DE]Attach. The invalid magic tells
1517 * srpc_handle_rpc to drop this RPC */
1521 if (!list_empty(&scd->scd_rpc_free)) {
1522 srpc = list_entry(scd->scd_rpc_free.next,
1523 struct srpc_server_rpc,
1525 list_del(&srpc->srpc_list);
1527 srpc_init_server_rpc(srpc, scd, buffer);
1528 list_add_tail(&srpc->srpc_list,
1529 &scd->scd_rpc_active);
1530 swi_schedule_workitem(&srpc->srpc_wi);
1532 list_add_tail(&buffer->buf_list,
1533 &scd->scd_buf_blocked);
1536 spin_unlock(&scd->scd_lock);
1538 spin_lock(&srpc_data.rpc_glock);
1539 srpc_data.rpc_counters.rpcs_rcvd++;
1540 spin_unlock(&srpc_data.rpc_glock);
1543 case SRPC_BULK_GET_RPLD:
1544 LASSERT (ev->type == LNET_EVENT_SEND ||
1545 ev->type == LNET_EVENT_REPLY ||
1546 ev->type == LNET_EVENT_UNLINK);
1549 break; /* wait for final event */
1551 case SRPC_BULK_PUT_SENT:
1552 if (ev->status == 0 && ev->type != LNET_EVENT_UNLINK) {
1553 spin_lock(&srpc_data.rpc_glock);
1555 if (rpcev->ev_type == SRPC_BULK_GET_RPLD)
1556 srpc_data.rpc_counters.bulk_get += ev->mlength;
1558 srpc_data.rpc_counters.bulk_put += ev->mlength;
1560 spin_unlock(&srpc_data.rpc_glock);
1562 case SRPC_REPLY_SENT:
1563 srpc = rpcev->ev_data;
1564 scd = srpc->srpc_scd;
1566 LASSERT(rpcev == &srpc->srpc_ev);
1568 spin_lock(&scd->scd_lock);
1570 rpcev->ev_fired = 1;
1571 rpcev->ev_status = (ev->type == LNET_EVENT_UNLINK) ?
1572 -EINTR : ev->status;
1573 swi_schedule_workitem(&srpc->srpc_wi);
1575 spin_unlock(&scd->scd_lock);
1586 memset(&srpc_data, 0, sizeof(struct smoketest_rpc));
1587 spin_lock_init(&srpc_data.rpc_glock);
1589 /* 1 second pause to avoid timestamp reuse */
1590 set_current_state(TASK_UNINTERRUPTIBLE);
1591 schedule_timeout(cfs_time_seconds(1));
1592 srpc_data.rpc_matchbits = ((__u64) cfs_time_current_sec()) << 48;
1594 srpc_data.rpc_state = SRPC_STATE_NONE;
1596 rc = LNetNIInit(LNET_PID_LUSTRE);
1598 CERROR ("LNetNIInit() has failed: %d\n", rc);
1602 srpc_data.rpc_state = SRPC_STATE_NI_INIT;
1604 LNetInvalidateHandle(&srpc_data.rpc_lnet_eq);
1605 rc = LNetEQAlloc(0, srpc_lnet_ev_handler, &srpc_data.rpc_lnet_eq);
1607 CERROR("LNetEQAlloc() has failed: %d\n", rc);
1611 rc = LNetSetLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1613 rc = LNetSetLazyPortal(SRPC_REQUEST_PORTAL);
1616 srpc_data.rpc_state = SRPC_STATE_EQ_INIT;
1624 srpc_data.rpc_state = SRPC_STATE_RUNNING;
1630 srpc_shutdown (void)
1636 state = srpc_data.rpc_state;
1637 srpc_data.rpc_state = SRPC_STATE_STOPPING;
1642 case SRPC_STATE_RUNNING:
1643 spin_lock(&srpc_data.rpc_glock);
1645 for (i = 0; i <= SRPC_SERVICE_MAX_ID; i++) {
1646 srpc_service_t *sv = srpc_data.rpc_services[i];
1648 LASSERTF (sv == NULL,
1649 "service not empty: id %d, name %s\n",
1653 spin_unlock(&srpc_data.rpc_glock);
1657 case SRPC_STATE_EQ_INIT:
1658 rc = LNetClearLazyPortal(SRPC_FRAMEWORK_REQUEST_PORTAL);
1659 rc = LNetClearLazyPortal(SRPC_REQUEST_PORTAL);
1661 rc = LNetEQFree(srpc_data.rpc_lnet_eq);
1662 LASSERT (rc == 0); /* the EQ should have no user by now */
1664 case SRPC_STATE_NI_INIT: